Research And Application Of The Learning Algorithm For Spiking Neural Networks

The development of Spiking Neural Networks is a further study and development of brain-like computing.In comparison with the traditional neural network mechanism,the spiking neural network simulates the biological neuron,imitating the structure of the biological neuron,and at the same time incorporates time information into the coding method.Therefore,the spiking neural network carries both space and time information and can express richer concepts.At present,Spike Neural Networks have relatively good algorithms and simple applications in image recognition,computer vision,and speech processing.However,due to the relatively short development history of spiking neural networks,scientific researchers need to explore more in theory and application.In the development of spiking neural networks,effective algorithms are the focus of research.Compared with the landmark convolutional neural network(CNN)and recurrent neural network(RNN)of traditional neural networks,pulsed neural networks require further attempts in the development of single-layer network algorithms and deep network.Currently,due to the lack of efficient and accurate learning algorithms,the development and application of spiking neural networks have encountered major bottlenecks.Based on the existing classic algorithms,this thesis has made the following innovative attempts:1.Based on the understanding of biological delay learning mechanism,optimize and innovate the existing classic algorithm Temporron.Firstly,the learning method of the delay mechanism is derived,and then based on the delay mechanism,the DWTP algorithm is proposed.Experimental results show that the DWTP algorithm,which can adjust the two parameters of weight and delay time,improves the learning efficiency by30% and has better robustness compared with the Tempotron algorithm which only relies on weight adjustment.2.As the Temporron algorithm belongs to the binary classification algorithm of single-layer neural network,even if the algorithms such as Multi-Tempotron are gradually optimized in the subsequent development process,it still has limitations in solving complex problems.Therefore,the algorithm optimization of multi-layer networks is an innovative direction.DWSBP algorithm is proposed by dynamically optimizing the delay time of Spike Prop algorithm.Experimental results show that DWSBP algorithm with delay mechanism regulation also has higher learning efficiency and stronger ability to resist noise,and learning efficiency is increased by 16%.3.Finally,considering the complexity of the environment in the practical problem,the image feature extraction can significantly improve the recognition efficiency of various algorithms.Therefore,CNN convolutional networks is used as the preprocessing method of feature extraction,and DWTP and DWSBP algorithms proposed in this thesis are fused as classifiers to form a recognition model.The experimental results show that this model is similar to the existing pulse neural network pattern recognition model.